Multi-instance image editing on LoMOE-Bench (test)
26Target ConsistencyLoMOE
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| LoMOE2026.02 | 26 | 0.067 | 0.266 | 0.546 | 0.09 | 0.834 | 29.4 | 98.96 | |
| FLUX μTscenario=multi-turn2026.02 | 25.71 | 0.074 | 0.282 | 0.55 | 0.15 | 0.873 | 28.27 | 94.27 | |
| LayerEdit2026.02 | 25.61 | 0.071 | 0.252 | 0.186 | 0.147 | 0.864 | 29.07 | 100 | |
| OursEfficient prompt encoding=true, IDAttn=true2026.02 | 25.6 | 0.073 | 0.277 | 0.574 | 0.099 | 0.919 | 29.08 | 89.06 | |
| FLUXprompting=end-to-end2026.02 | 24.71 | 0.078 | 0.266 | -0.059 | 0.206 | 0.83 | 27.58 | 94.79 | |
| FLUX w/ v.c.visual cues=drawing boxes2026.02 | 24.49 | 0.072 | 0.267 | 0.265 | 0.17 | 0.893 | 27.6 | 92.71 |